SafePost

Inspiration

The inspiration for SafePost came from the growing concern about privacy in our AI-driven world. As AI becomes more integrated into our daily lives, so do the risks of data leakage and identity theft. We wanted to build a solution that empowers users to protect their personal information while still enjoying the freedom of social media. Our goal was to tackle the hackathon's core challenge: using AI to protect user privacy and enhance the privacy of AI systems themselves.

What it does

SafePost is your personal privacy guardian for social media. It's a mobile application that uses AI to scan and protect your images and text before you share them.

Image Analysis: It automatically detects sensitive visual information like faces, documents, license plates, and addresses.
Text Analysis: It uses a lightweight LLM to scan your captions and notes for personal details such as travel plans, phone numbers, or financial information.
Privacy-First Protection: SafePost provides a real-time risk assessment and offers tools like AI blurring for images to help you share safely without giving away too much.

How we built it

We built SafePost as a completely on-device application to ensure maximum privacy. This means your data never leaves your phone to be processed on a cloud server.

On-Device AI: We used lightweight models like MobileNetV3 for object detection, an Executorch OCR model for text recognition, and SmolLM2 for text analysis. We pulled these models once from HuggingFace and run them entirely on the device.
Technology Stack: The application was developed with React Native and Expo for cross-platform compatibility. We used Android Studio and XCode for Android and IOS development, with Node.js and Metro Bundler handling the project's dependencies and packaging.
Optimized Performance: To provide a seamless user experience, we designed the image and text processing pipelines to run simultaneously, reducing latency and allowing you to start typing your caption while the image is still being scanned.

Challenges we ran into

Our biggest challenge was balancing model size with performance and accuracy. Running AI models locally on a mobile device required us to find lightweight models that could still reliably detect various types of PII without draining the phone's battery. We spent a lot of time optimizing the application to ensure a smooth, responsive experience.

Accomplishments that we're proud of

We are most proud of building a fully functional privacy tool that operates entirely on the edge. By keeping all processing local, we created a truly private solution that meets both of the hackathon's core objectives. We successfully managed the technical complexities of integrating and running multiple AI models on a mobile device in a way that is both efficient and intuitive for the user.

What we learned

We learned that it's completely possible to create powerful, privacy-focused AI applications that don't rely on cloud computing. This project taught us the importance of optimizing for on-device performance and the value of lightweight, efficient models. We also gained a deeper understanding of how to design a user experience that makes complex AI processes simple and approachable for everyone.

What's next for SafePost

We have big plans for SafePost's future:

Video Processing: We want to expand our detection capabilities to analyze videos and livestreams.
Advanced Editing: We plan to give users more control over image censoring and provide refined text suggestions from the LLM.
Social App Integration: The ultimate goal is to integrate SafePost directly into popular social media apps as a plugin, offering real-time privacy protection right where you need it.